[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-80841":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":14,"stars7d":13,"stars30d":13,"stars90d":15,"forks30d":15,"starsTrendScore":16,"compositeScore":17,"rankGlobal":10,"rankLanguage":10,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":21,"hasPages":19,"topics":22,"createdAt":10,"pushedAt":10,"updatedAt":23,"readmeContent":24,"aiSummary":25,"trendingCount":15,"starSnapshotCount":15,"syncStatus":26,"lastSyncTime":27,"discoverSource":28},80841,"ComfyUI-SPEED","ruwwww\u002FComfyUI-SPEED","ruwwww","ComfyUI Unnofficial Implementation of Spectral Progressive Diffusion for Efficient Image and Video Generation for Anima","",null,"Python",43,7,1,0,3,2.71,"MIT License",false,"main",true,[],"2026-06-12 02:04:07","# ComfyUI-SPEED\n\n> Warning: This repository is \"vibecoded\" — it may contain experimental code, unconventional styles, or project-specific shortcuts.\n\nThis repository provides a ComfyUI custom node that integrates the SPEED (Spectral Progressive Diffusion) idea for faster sampling. SPEED is a method for progressively growing image resolution during diffusion denoising to reduce computation while preserving visual quality. (only tested for anima model)\n\nKey references\n\n- Original project page: https:\u002F\u002Fhowardxiao.ca\u002Fspeed\u002F\n- Paper (PDF): https:\u002F\u002Fhowardxiao.ca\u002Fspeed\u002Fpaper\u002Fpaper.pdf\n- arXiv: https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.18736\n\nSummary\n\nWorkflow\n\n![Workflow](images\u002Fworkflow.png)\n\nSpeed comparison (Anima, using the current default input config)\n\n| SPEED sampler (this node)                                              | Baseline (euler sampler)                                              |\n| ---------------------------------------------------------------------- | --------------------------------------------------------------------- |\n| `base_sampler=euler`\u003Cbr>`start_scale=0.50`\u003Cbr>`mid_scale=0.75`\u003Cbr>`transition_1=0.80`\u003Cbr>`transition_2=0.60`\u003Cbr>`taper=8`\u003Cbr>\u003Cbr>![SPEED](images\u002Fanima_speed.png)\u003Cbr>\u003Cbr>**14.55s**\u003Cbr>**1.82× faster** | ![Original](images\u002Fanima_original.png)\u003Cbr>\u003Cbr>**26.51s**\u003Cbr>**1.00×** |\n\nNotes\n- **Artifacts:** This implementation can produce visible artifacts on some outputs; results may vary by model and prompt. Inspect the example images above for a representative comparison. `taper` (default `8`) crossfades the DCT seam to reduce ringing at transitions; set `taper=0` for the original hard-truncation behaviour.\n- **Torch compile:** Compiling with `torch.compile` did not improve performance for this implementation and in our tests made sampling slower than running without it. It may be possible for others to make the node work with `torch.compile`, but this remains a known \u002F open issue.\n- Spectral Progressive Diffusion (SPEED) progressively increases resolution and injects higher-frequency components along the denoising trajectory, enabling training-free acceleration and a light fine-tuning recipe.\n- Personally i recommend using `transition_1` value of `0.8` and `transition_2` value of `0.7`. at 1.4x speed up for anima after many tries.\n\nUsage\n\n- Connect the `Sampler SPEED (Spectral Progressive)` output to `SamplerCustomAdvanced` like any other custom node in ComfyUI.\n- Place this folder under your ComfyUI `custom_nodes` directory, then restart ComfyUI.\n\nInputs\n\n- `base_sampler`: which underlying ODE solver to use. Any sampler in `comfy.k_diffusion.sampling` is supported (`euler`, `euler_ancestral`, `heun`, `dpmpp_2m`, `dpmpp_3m_sde`, `er_sde`, `res_multistep`, `uni_pc`, `ipndm`, …). `euler` is the original SPEED behaviour. Multistep solvers reset to first-order at each SPEED transition (solver state across a resolution change is invalid anyway).\n- `start_scale`: the first resolution fraction. `0.5` means the sampler starts at half resolution, so it does the earliest denoising work on a smaller image.\n- `mid_scale`: the second resolution fraction. `0.75` means it expands to 75% resolution before going to full size.\n- `transition_1`: Controls the jump to mid-scale; lower values delay the expansion, forcing the model to spend more steps inside the low starting resolution.\n- `transition_2`: Controls the jump to full-scale; setting this too low delays the final expansion, forcing the model to bake fine textures into a mid scale latent which stretches into blocky artifacts.\n- `taper`: DCT-bin width of the cosine crossfade at the preserved↔noise boundary. `0` reproduces the original hard truncation (more ringing); higher values smooth the seam at the cost of slightly more low-pass. Default `8`.\n\nCredits & license\n\n- This implementation links to and builds on the ideas from \"Spectral Progressive Diffusion for Efficient Image and Video Generation\" by Howard Xiao, Brian Chao, Lior Yariv, and Gordon Wetzstein. Please see the original project page and paper for full details, authorship, and license information.\n- Thanks to [acidmiku](https:\u002F\u002Fgithub.com\u002Facidmiku) for contributing the pluggable sampler dispatch, cosine taper crossfade, and DCT matrix cache.\n\nBibTeX\n\n```\n@article{xiao2026spectral,\n  author    = {Xiao, Howard and Chao, Brian and Yariv, Lior and Wetzstein, Gordon},\n  title     = {Spectral Progressive Diffusion for Efficient Image and Video Generation},\n  year      = {2026},\n}\n```\n","该项目提供了一个ComfyUI自定义节点，集成了SPECTRAL PROGRESSIVE DIFFUSION (SPEED)技术以加速图像和视频生成。其核心功能是通过在扩散去噪过程中逐步提高图像分辨率来减少计算量同时保持视觉质量，适用于Anima模型。技术特点包括支持多种基础采样器（如euler、heun等），以及可调节的参数如起始分辨率比例、过渡点等，用于优化速度与质量之间的平衡。适合需要高效生成高质量动画场景使用。",2,"2026-06-11 04:02:31","CREATED_QUERY"]